We read and discuss
EfficientDet: Scalable and Efficient Object Detection by Mingxing Tan, Ruoming Pang & Quoc V. Le https://arxiv.org/pdf/1911.09070.pdf

meetup recordings on youtube

Notes:

  • Many statements are based on intuition
  • Architecture: Bi-directional feature pyramid network + skip connections + learned weighting + compound scaling
  • Training: Focal loss, Swish activation
  • Some interesting ablation studies regarding the scale jitter for longer training schedules
  • For the next time: CLIP
  • Computer Vision Meetup Berlin needs an organizer